Where the Model Runs

Up to this chapter everything happened in a browser: you opened the service, typed a question, attached a file. For the tasks in the earlier chapters that is enough.

There are other ways to open a chat. The way you choose decides two things: where your text goes and which files the service can reach.

Browser and phone

You open a website or an app on your phone. Everything else happens on someone else's servers.

The question travels over the network to the company that makes the service. It is processed there, and the answer comes back to you. That is how ChatGPT, Claude, Gemini, and Grok work.

The service cannot reach files on your computer. Whatever you want to show it, you bring yourself by dragging the file into the conversation window.

A program on your computer

The same services have a program you install on your computer. The chat window in it is the same, but it works with files differently.

In the program you can create a project and link it to folders on your computer. The program creates files right in those folders or works with source data that is already there. You do not have to bring files in one at a time.

The model still sits in the cloud. The files the program read, or fragments of them, go to the company's servers along with the description of the task. The text leaves your computer just as it does from a browser.

A model on your own computer

The third option is for experienced users. It is a specialized setup, and it is described here so you know it exists.

The model itself is installed on your computer or on a company server. That is called a local model. Neither the request nor the file goes anywhere.

You install it with a separate program. Ollama and LM Studio download the model you pick onto the computer, and it works as well as the machine allows. The more memory and the stronger the graphics card, the larger the model you can run and the faster it answers.

On an average computer with no dedicated graphics card, a model answers far more slowly than a cloud service. For an experiment that is fine, and it is worth trying. For everyday work it is a poor fit.

You pay in quality: a model that fits on a laptop is weaker than a cloud one and can do less.

Companies run models in-house to develop them further by training on their own data and to use them in internal systems. The other reason is data that may not leave the company.

That requires serious infrastructure. Models like these run on powerful, expensive servers with professional graphics cards.

Corporate access

Companies buy corporate access to a cloud service.

The data still goes out, but under a contract. The contract forbids using it for training and sets retention periods.

The company decides this, not the employee. What is allowed at yours is written in its policies.

Access to a folder is wider than access to a file

When you talk to a model through a browser and attach one file, the model gets that file. When you attach a folder, you are attaching everything inside it.